Tesla’s FSD demonstrates impressive superhuman reflexes, yet its failure to detect simple physical barriers like chains proves that raw speed cannot yet replace nuanced human perception. The ongoing struggle to balance aggressive safety with social predictability remains the system's most significant hurdle.
Deep Dive
Prerequisite Knowledge
- No data available.
Where to go next
- No data available.
Deep Dive
Finding the limits of Tesla Self-Driving
Added:Oh yeah.
>> Maybe go really low.
>> Oh.
>> Oh no. Wow. That's crazy. All right.
How's it going, guys? This is a FSD test video. We got David Moss sitting up there with a flagger outfit, stop sign, slow sign. We're going to do a bunch of testing today. Kind of a little bit of everything. So, let's see what FSD version 14.3.4 is capable of. So, we've set up a test to see exactly how full self-driving handles a human flagger.
That's David Moss up ahead in a vest holding a stop sign. There's a huge difference between a regular street sign and this. At a standard intersection, the car stops, yields, and then proceeds. But here, the system has to understand the context. It doesn't just get a go like usual. It has to recognize David as a flagger and actually wait until he flips the sign to yield. And you can see that's exactly what full self-driving does here. Once the sign says yield, the car proceeds. David does quickly turn the stop sign, but it's a little bit too late and the car keeps going, which I think is okay here. And now we wanted to test the reaction time of full self-driving version 14.3.
Okay, I'm coming fast.
Oh yeah, good job.
So, I want to kind of break down everything here and without getting too technical, explain a few things. This timeline that I'm in is 60 frames per second and everything that I've recorded. All the videos are also in 59.94 or 60 frames per second. What that tells us is it allows us to frame by frame analyze something and then give basically a reaction speed for FSD. So you can see here we're driving. The sign is on slow and then I have a marker here. But basically right we can say right here is the first frame where like it begins to look like a stop sign. So from here it's 1 2 3 4 5 6 7 8 9 10 11 12 13 frames. And you can see now the planner here is slowing down aggressively. This is 13 frames later and the car is now reacting to the stop sign and slowing down. So if you put that information into Gemini and I actually must have miscounted it. It's 13 or 14 frames. Very close. So you can see here are all the details. This is from Gemini giving me basically a readout where if you translate those 14 frames and convert it to a reaction time, it's better than a race car driver, better than any human ever could be. And at the bottom here kind of sums it up as saying, "The software is operating at the absolute limit of biological human capability. At 234 milliseconds, the autonomous system is reacting roughly three times faster than an average attentive human driver, operating squarely in the realm of an elite race car driver anticipating a corner." And here's just another clip showing how fast the car is reacting to the slow sign being spun to stop. Again, something that may not come across on camera is just how fast we are going and how quickly the car comes to an aggressive but safe stop in this situation.
Oh, that was good.
How far can you back me up? So, this is just kind of showing off something that maybe not everybody realizes, but full self-driving will do everything it can to get out of the way of a person. So, it does the same behavior if a car is backing up into it. And right here, we can see David kind of is bullying the car backwards. It's trying to turn on the signal and go around him. But above all else, full self-driving is doing everything it can to not come in contact with the person. So you can see here David has full confidence standing in front of this car that nothing is going to happen and all it wants to do is get away from him and do so safely. Next up is our version of the thin chain test.
Something you've probably seen Chuck Cook do on his channel. But instead of a chain, David is holding up a strip of duct tape right across the road to see if the vision system can actually pick it up. I've got to be honest, I was completely surprised by the results on this one. So before we roll the clip, do me a favor, pause the video right now and drop a comment below. Do you think FSD is going to stop for the tape? Let's find out. Okay, so what we're doing now is we're testing a thin chain. Uh we have a piece of duct tape across. Dave is going to hold it and we'll just try various things either like picking it up, leaving on the ground um and see how the car does with visualizing it and stopping for it. Uh >> I'll just let it I'll just let it hit but not have much tension so it doesn't >> Yeah. Yeah. Yeah. Let it let it hit it.
>> Maybe keep it like a teeny bit lower like where it ch like Yeah. Like that.
That feels like it should see it there.
Oh, yeah. It's perfect. Okay. Oh, >> did it?
>> You got to adjust the 360.
>> Yeah. Yeah. Huh. I'll put it in standard. Not that I think that that should matter. And I'll hit start.
Okay. So, we're going coming around the corner. I can see it.
I can see it. I can see it. I can see it. I can see it. the car. Okay, much better that time >> cuz now it's treating it like a a blocked road as it should.
>> Weird.
>> All right, let's try it from the left again.
>> Very visible to me.
Okay, so why this time? Here, let me go.
>> It went It went past it though.
I mean like you were Oh, I see what you're saying. Yeah, >> you were to the windshield.
>> Okay, starting.
>> I think like that like that's in motion and everything like that should be plenty easy to see.
>> Stop. Stop. Stop.
>> Oh, it would have knocked over the camera. Okay. Well, that's a fail.
>> I was waving it, too. Yeah, like it was moving. It was That was so visible.
>> Oh, I'll do it a little higher this time, but >> starting FSD.
>> Okay, come on. It has to. Oh. Oh. Oh, yeah. It sees it when it's in line with this camera much more, but it still will hit it. Crazy.
Well, I'd say that's a fail, >> right?
>> All right. Do one more. Yeah, do one more, though. interrupting the video for two seconds just to tell you about today's sponsor, Sentry Pro. The official Tesla app actually only alerts you if your alarm is going off, where Sentry Pro will let you know if there's somebody approaching the vehicle or touching the vehicle. Those are instant realtime alerts that get pushed directly to your phone. And with the Sentry Pro app, you can actually program the Tesla to flash the lights, honk the horn, and do more, turning your car from being just this passive monitoring thing to an active deterrence. This is a feature that I genuinely assumed came with the car when I first got it and was surprised when I wasn't getting notifications and I'd come back to the car, see all the sentry alerts, and not have anything on my phone. And the app itself is great. It's uh developed by an official Tesla software developer. It's using Fleet API. There's no data tracking, no data reselling, and very minimal data access required to use.
Another thing that's great about this is it works without premium connectivity.
So, anybody that has a car without premium connectivity, this works for you. works on all Tesla models and it's constantly getting updates and improvements to the app for more and more functionality. And so right now there's a free trial as well as 20% off a lifetime subscription. You can use my code that's going to be linked in the description below as well as I'll put it here on the screen right now. If you have any questions at all, let me know.
I use this app. I enjoy it. I think it's great and I think you'll like it, too.
>> Oh yeah, that's a good that's a classic chain across the road. I think it's going to hit it.
Oh, I think we broke it, didn't we?
>> Yeah.
>> So, how's it going? Hey, >> wait. What do you got?
>> Chain. Duct tape chain. And we failed.
>> I feel like that's I mean, it's very visible to me from the car.
>> Yeah.
>> And then it stopped when it was higher up. I honestly thought it would see it better when you had it lower like that.
>> Maybe if we weren't such broke pieces of We had a front bumper camera, this would be different.
>> That would actually be Wow. We're both terrible because we should have the front bumper one to test that. You're right.
>> So, that was several tests with effectively recreating a thin chain and surprisingly full self-driving did a pretty poor job here. It did catch it a couple times, but not enough and it shouldn't be missing something as thick and as visible as duct tape. I know duct tape across the road is not a very like normal thing, but it does emulate the idea of a chain and I do think anything like that the car should be able to see and react to. So, a little bit disappointed to see it fail this test as many times as it did. But, it is nice cuz now that's something we can test again in the future as new versions of FSD do come out. Driving past cars can be stressful, especially when a door suddenly swings wide open. Today, we're testing exactly how FSD handles this.
First, take a look at this clip. This is showing FSD driving by a car where the door doesn't open, just to give you an idea of how it should look. But in these next few examples, watch how quickly the system adjusts. The moment that the door opens, FSD instantly reacts. The camera spot the hazard instantly, and the car rapidly swerves left to avoid collision.
It is reacting faster than most human drivers ever could. But this raises a huge debate. Does FSD swerve too much?
While avoiding the door is undeniably safe, making a sudden, aggressive move into the next lane can make other drivers on the road really uncomfortable. It's a tough balance between pure safety and predictability.
The raw reaction speed is impressive throughout all of these tests, but smoothing out that evasion is definitely something that could be refined in a later version of FSD version 14. Let me know what you guys think in the comments. Does that aggressive swerve make you nervous, or is it acting like you'd want the car to?
>> All right, David, what are we up to?
>> We got some more shenanigans going on.
So, we're on the new camera preview screen. Shout out to Owen Sparks for teaching us this new trick, but we're going to go down the road and then and then bam.
>> And then block both those cameras. We have kind of like a funny little setup here, but it'll >> with this new grid view. And since Devon doesn't have a front bumper camera like myself, we don't even get that feed.
>> I know. But what do you think's going to happen? I genuinely don't know.
>> I think it'll just go red hands.
>> Red hands and stop. Yeah. Right.
>> I wonder if it'll fix itself, though.
I wonder if you can leave it for a long time and then fix it and if it'll go.
>> We'll try.
>> We'll try. We'll see.
>> Okay. So, we're driving normally whenever you want.
Okay.
System error.
Oh, back on. Oh, pretty good.
I mean, it's kind of what I expected.
Maybe go really low.
>> Oh, you can't.
>> Oh. Oh, I forgot about that. One sec. One sec.
All right.
All right. The auto wipers are really >> really crucial.
>> Really crucial here. Okay. We're going to try that again. I'm going to immediately turn off auto wipers. Okay.
So, this is what happens. Blocked.
Blinded.
No red hands yet. I mean, that's fully blocked.
Bull. No. Wow. That's crazy. That is full full full block.
>> Very like that is the camera preview and in front of it just thinks there's a wall. How long is it going to back up?
>> The wall. No hands.
>> What?
>> This is hilarious.
>> Oh, crazy. Is it just going to park itself? I think that's what it should do right now cuz it's so uncertain.
No.
Oh my god.
It's trying to go forward and then every time it tries there's a wall so it goes backwards because there's a huge thing in front of it. The Insta 360 will be showing it.
Oh. Oh. Oh.
Is it going to go up on the curb?
>> It's not even backing up good.
>> No, it's doing a really bad job.
What?
>> It's too It's too perfect.
>> Oh no.
At what point's it going to stop?
We're up on the curb.
>> It's only on the sidewalk.
>> It's on the sidewalk.
>> Okay.
>> You think you think change it >> once it goes back to drive?
>> Really? No. No. I I want to see like >> play this out.
>> Well, I don't know. Do you think it's going to hit something?
>> You're Subaru. No, look. I mean, it's >> I'd rather it stop. I wanted to understand that it should park itself.
>> It's going right for that, but we're on >> We're fully on the curb.
>> All right. I mean, it's your car. Yeah.
Yeah. Let's Let's let it I want to see if it'll come in contact with something, which it almost is. Oh my god, it almost is.
>> It couldn't be any closer.
>> Okay, I think we're Yeah, we're just about touching that hedge.
>> Oh, we're touching the edge.
Oh, it parked. Okay.
I mean I I mean at least I didn't mess up the hedge, but >> yeah.
>> Show the show the camera we're like literally on.
>> Okay, let's see. Let's see where we parked and what the camera is. So that's the camera and this is where we've ended up parked.
I couldn't couldn't be closer.
>> Crazy. I mean, it didn't hit it. I'll give it that.
>> Is it touching?
>> No, it's like I I It's like >> rubbing.
>> Rubbing >> crazy.
>> All right, this brings us to the final test of the day. and honestly the one I've been looking forward to the most.
Meet Greg. He's an inflatable dummy that David and I dress up to look like a person. The plan here is pretty simple.
We're going to have David throw Greg right into the middle of the road as the Tesla approaches just to see exactly how FSD handles the sudden pedestrian hazard.
>> Greg.
Wow.
That was pretty good, eh?
>> David tosses Greg out with a decent amount of warning and watch the reaction here.
>> Wow.
Okay, wait, wait, wait, wait, wait, wait. Walk away.
We're not too worried about Greg, but we're not running him over. Okay, that's good.
Immediately, the car spots the hazard.
But what stood out to us throughout all of these runs wasn't just that it stopped. It was how incredibly fast it reacted and the aggressive evasive maneuvers it took to ensure it didn't come anywhere near Greg. It didn't just slam on the brakes in a straight line and hope for the best. It instantly calculated an escape route and physically moved out of the way.
Okay.
One sec. Just see what it does. See what it does.
Poor Greg.
What do they do to my boy?
Okay, it did that. It did that. We even tried with David hiding in a bush just around a blind corner. As soon as the car cleared the turn, Greg goes flying into lane. And even then, FSD spotted him instantly, swerved, and kept Greg completely safe.
Okay. See what it does here.
I'm looking at the drone. That's why I'm looking this way.
FSD doesn't like some man could have a heart attack on the road and the car is like just like I'm out of here. No matter what we tried or how late we threw, we absolutely could not get full self-driving to hit the dummy. And that wraps up our testing for the day. And honestly, I'm walking away pretty impressed. The progress we are seeing with the software is undeniable. The sheer safety of the system has gotten to a point where it's hard to comprehend just how much it's analyzing in a fraction of a second. The reaction time improvements we are seeing here in version 14 are absolutely remarkable. It really feels like the speed at which FSD's overall abilities are increasing is just getting faster and faster with every single update. This is actually my first time doing a fully narrated format like this, stepping behind the mic just to break down the footage instead of just letting it play out. I really hope you guys enjoyed this style and found it helpful in understanding what the system is actually doing behind the scenes. If you did enjoy it, please leave a like on the video. It helps me out more than you know. And drop a comment below letting me know what you think the best test we did today was. Make sure you're subscribed because we have a lot more testing coming up very soon. Thanks for watching and I'll see you in the next one.
Related Videos

Setting up a curved screen with Immersive Calibration Pro 4 and multiple cameras (P3D v4)
FlyerOneZero
23K views•2019-07-21

Robot Learning with Sparsity and Scarcity
allenai
379 views•2025-10-14

Jorge Mendez-Mendez: Unlocking Lifelong Robot Learning With Modularity (2023-10-05)
umassmlfl
237 views•2024-01-06

Northwestern’s MS in Robotics: Student Robotics Projects, 2023
NorthwesternEngineering
1K views•2024-05-31

"Perfect" Turns: Turning by the Gyro - FIRST LEGO League (FLL) SPIKE Prime + EV3 RePlay Programming
ZacharyTrautwein
94K views•2020-10-02

Gorkem Secer: TSLIP-based Deadbeat Running Control of Bipedal Robot ATRIAS
DynamicWalking-wv6qm
298 views•2018-06-22

Self-Driving Cars Need Lessons On Human Drivers | Maddie About Science
skunkbear
26K views•2018-08-21

Milrem Robotics’ THeMIS UGVs used in a live-fire manned-unmanned teaming exercise
MilremRobotics
99K views•2021-05-20
Trending

Playstation NO DISC/NO BUY Fight Is Over...
DavidJaffeGames
4K views•2026-07-23

Steam and Xbox Just Dropped The Hammer On PlayStation
OhNoItsAlexx
9K views•2026-07-23

Americans Confused in Australia for 17 Minutes Straight
IWrocker
17K views•2026-07-23

SuperBike Factory Has Gone... What's Next for the Motorcycle Industry?
thatbikersimon
11K views•2026-07-22